keras剪枝-量化-推理

待我称王封你为后i 2023-01-10 10:19 112阅读 0赞

tensorflow提供了一个优化工具tensorflow\_model\_optimization ，专门针对keras进行模型优化

主要可以进行剪枝、量化和权重聚类

这里主要使用前面两个

数据集使用以前的文章：[mnn模型从训练-转换-预测][mnn_-_-]

具体训练代码如下

注意：使用之前需要手动安装tensorflow\_model\_optimization，使用pip install tensorflow\_model\_optimization就行

import tempfile
    import os
    import tensorflow as tf
    import numpy as np
    from tensorflow import keras
    from tensorflow.keras import layers
    from tensorflow.keras.models import Sequential
    import tensorflow_model_optimization as tfmot
    
    
    
    batch_size = 2
    img_height = 180
    img_width = 180
    num_classes = 5
    epochs = 50
    validation_split=0.2
    data_dir='flower_photos'
    
    #数据集准备
    train_ds = tf.keras.preprocessing.image_dataset_from_directory(
      data_dir,
      validation_split=validation_split,
      subset="training",
      seed=123,
      image_size=(img_height, img_width),
      batch_size=batch_size)
     
    val_ds = tf.keras.preprocessing.image_dataset_from_directory(
      data_dir,
      validation_split=validation_split,
      subset="validation",
      seed=123,
      image_size=(img_height, img_width),
      batch_size=batch_size)
    
    AUTOTUNE = tf.data.experimental.AUTOTUNE
    train_ds = train_ds.cache().shuffle(1000).prefetch(buffer_size=AUTOTUNE)
    val_ds = val_ds.cache().prefetch(buffer_size=AUTOTUNE)
    
    
    model = keras.Sequential([
      keras.layers.InputLayer(input_shape=(img_height, img_height,3)),
      keras.layers.Reshape(target_shape=(img_height, img_height, 3)),
      layers.Conv2D(16, 3, padding='same', activation='relu'),
      layers.MaxPooling2D(),
      layers.Conv2D(32, 3, padding='same', activation='relu'),
      layers.MaxPooling2D(),
      layers.Conv2D(64, 3, padding='same', activation='relu'),
      layers.MaxPooling2D(),
      layers.Dropout(0.2),
      layers.Flatten(),
      layers.Dense(128, activation='relu'),
      layers.Dense(num_classes)
    ])
    
    model.compile(optimizer='adam',
                  loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                  metrics=['accuracy'])
     
    print(model.summary())
    model.fit(
      train_ds,
      validation_data=val_ds,
      epochs=epochs
    )
    
    tf.keras.models.save_model(model, 'baseline_model.h5', include_optimizer=False)
    converter = tf.lite.TFLiteConverter.from_keras_model(model)
    tflite_model = converter.convert()
    open("baseline_model.tflite", "wb").write(tflite_model)
    
    
    #开始剪枝
    print("start pruning")
    prune_low_magnitude = tfmot.sparsity.keras.prune_low_magnitude
    num_images =3670
    end_step = np.ceil(num_images / batch_size).astype(np.int32) * epochs
    pruning_params = {
          'pruning_schedule': tfmot.sparsity.keras.PolynomialDecay(initial_sparsity=0.50,
                                                                   final_sparsity=0.80,
                                                                   begin_step=0,
                                                                   end_step=end_step)
    }
    
    model_for_pruning = prune_low_magnitude(model, **pruning_params)
    model_for_pruning.compile(optimizer='adam',
                  loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                  metrics=['accuracy'])
    model_for_pruning.summary()
    logdir = tempfile.mkdtemp()
    callbacks = [
      tfmot.sparsity.keras.UpdatePruningStep(),
      tfmot.sparsity.keras.PruningSummaries(log_dir=logdir),
    ]
    
    model_for_pruning.fit(train_ds, 
                      batch_size=batch_size, epochs=5, validation_data=val_ds,
                      callbacks=callbacks)
    model_for_export = tfmot.sparsity.keras.strip_pruning(model_for_pruning)
    #开始量化
    print("start quantize")
    quantize_model = tfmot.quantization.keras.quantize_model
    q_aware_model = quantize_model(model_for_export)
    q_aware_model.compile(optimizer='adam',
                  loss=tf.keras.losses.SparseCategoricalCrossentropy(from_logits=True),
                  metrics=['accuracy'])
    
    q_aware_model.summary()
    q_aware_model.fit(train_ds,
                      batch_size=batch_size, epochs=5, validation_data=val_ds)
    converter = tf.lite.TFLiteConverter.from_keras_model(q_aware_model)
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    quantized_and_pruned_tflite_model = converter.convert()
    quantized_and_pruned_tflite_file ='pruned_and_quantized.tflite'
    with open(quantized_and_pruned_tflite_file, 'wb') as f:
      f.write(quantized_and_pruned_tflite_model)

运行结束后，我们看一下模型文件：

![20210123171122714.png][]

可以看到文件确实压缩了不少了，4倍左右

接下来试一下，推理速度

优化过的模型文件推理如下

import tensorflow as tf
    import cv2
    import numpy as np
    import time
    
    start=time.time()
    image = cv2.imread('397.jpg')
    image=cv2.resize(image,(180,180))
    image=image[np.newaxis,:,:,:].astype(np.float32)
    print(image.shape)
    interpreter = tf.lite.Interpreter(model_path='pruned_and_quantized.tflite')
    interpreter.allocate_tensors()
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    for _ in range(10):
        interpreter.set_tensor(input_details[0]['index'],image)
        interpreter.invoke()
        output_data = interpreter.get_tensor(output_details[0]['index'])
    print(output_data)
    print('avg infer time is %.6f s'%((time.time()-start)/10.0))

运行结果：

![20210123165345699.png][]

原始模型推理：

import tensorflow as tf
    import cv2
    import numpy as np
    import time
    
    start=time.time()
    image = cv2.imread('397.jpg')
    image=cv2.resize(image,(180,180))
    image=image[np.newaxis,:,:,:].astype(np.float32)
    print(image.shape)
    interpreter = tf.lite.Interpreter(model_path='baseline_model.tflite')
    interpreter.allocate_tensors()
    input_details = interpreter.get_input_details()
    output_details = interpreter.get_output_details()
    for _ in range(10):
        interpreter.set_tensor(input_details[0]['index'],image)
        interpreter.invoke()
        output_data = interpreter.get_tensor(output_details[0]['index'])
    print(output_data)
    print('avg infer time is %.6f s'%((time.time()-start)/10.0))

运行结果：

![20210123171158412.png][]

出乎意料的是，模型文件虽然变小了，但是速度居然还慢了，还慢了几十倍

这里是使用的ubuntu，也试过了win10上面推理差距更大，慢更多倍

反正就是一句话，优化后居然慢了。。。。

只有是找不到tflite的benchmark工具，只能使用这样的方式进行测试，也许时间不靠谱

[mnn_-_-]: https://blog.csdn.net/zhou_438/article/details/111563108
[20210123171122714.png]: /images/20221119/0534249fdba144f5b999a4277643f710.png
[20210123165345699.png]: /images/20221119/117e8015e0d3408693336cc87d765655.png
[20210123171158412.png]: /images/20221119/fcb975e8c96544b1973e9401eced18eb.png